Definition drives design: Disability models and mechanisms of bias in AI technologies

How AI systems are designed depends on how we define the problem: The paper demonstrates that different conceptual models of disability (medical, social, and biopsychosocial) lead to fundamentally different design choices in AI systems—from what data gets collected to how the technology is ultimately used—meaning bias isn't just a technical glitch but emerges from foundational definitional choices made early in development.

Bias accumulates through multiple "innocent" decisions: Rather than stemming from a single flawed algorithm, bias in AI systems targeting disabled people emerges from a cascade of seemingly neutral design decisions across problem formulation, data selection, use cases, and operational elements, making it harder to identify and fix where it begins.

Disabled people are largely absent from the design process: The paper emphasizes that lack of transparency and disabled participation throughout AI development amplifies potential harms, suggesting that meaningful inclusion of disabled people in design decisions could help catch biased assumptions before they become embedded in systems.

Current bias mitigation efforts may be insufficient: While significant work exists on analyzing and mitigating algorithmic bias, the paper argues these efforts don't adequately address the broader mechanisms of how bias emerges, indicating that technical fixes alone won't solve the problem without addressing upstream design decisions.

Definitions matter for real-world consequences: How disability is defined in AI systems used for healthcare, employment, benefits, and government policy directly determines who gets helped or harmed, making definitional choices a critical equity issue rather than merely an academic one.